Re: An unusual Performance approach using Synthetic registers

[dewar@gnat.com (Robert Dewar)]

> 1) I don't fully understand why synthregs aren't a common area rather
>    than part of stack frames.  A common area _adds_ code to
>    save/restore synthregs -- but it also increases the number and
>    frequency of references to synthregs.  I don't think L1 is the only
>    cache that can be used better by synthregs.

But if you put the SR's in a global area, then indeed they WILL require
6 byte instructions for their access, and that can greatly increase
pressure on the instruction cache, and will likely slow things down
greatly. 

Once again, you really can not discuss this in the abstract without looking
at the actual code sequences.

If you decide to allocate a base register, as was proposed at one point,
for the synth registers, then you are losing one of your real registers,
and that is a huge hit, you won't begin to buy that back.

With a change in the ABI, and OS, you could use FS or GS as the base
register, but that still does not help much, since the code wold
still be worse than normal access to local variables.

Remember again that the code to access SR's can be no better than the
code we generate right now for all accesses to local variables, function
arguments in memory, and spilled registers.